Track disposition of the #185 explain-lab probe framework (.scratch/)

[mobileskyfi]

Parent: #185 (phase-0 lab, now complete) · #90 (explain umbrella).

Phase 0 (#185) is answered and the ratification PR is up. The lab's probe framework currently lives only in .scratch/ (gitignored, non-authoritative, uncommitted). Per the constitution the probes are throwaway and the findings are the deliverable — but a good part of this framework is reusable when explain is actually implemented, and right now its only durable record is the sha256 hashes embedded in the #185 exit-artifact comments. This issue tracks the disposition so the work is neither silently lost nor smuggled into mainline.

What exists

.scratch/explain-lab-* — 76 files, ~16 MB (56 .ts probes/SUTs + JSON fixtures/captures). The two 7.5 MB files are the Q13 per-character highlight streams (explain-lab-q13-streams.v7.23.2.json / v7.24rc2.json), captured from CHR; everything else is small.

Rough shape:

• The offline SUT (a schema-free parser, written before scoring): explain-lab-segmenter.ts, -pathresolve.ts, -symbols.ts, -verbsplit.ts. This is the closest thing to a reference for the eventual grown canonicalizeExecuteCommand — reference only, it is throwaway and has the known Q17 DoS + Q14 fabrication defects on purpose.
• The oracle side: -q0-projector.ts (IL→{path,verb,args,blocks}), -cmdtree.ts, -cliref.ts, plus the frozen gold set -q0-gold-rev3.json.
• The frozen partition: -partition.ts / -partition.json (dev/holdout split, the method precondition for Q1–Q4/Q13).
• Per-question scorers + corner suites: -q{1,2,3,4,6,8,10,11,13,14,15,16,17}-*.ts.

Reusable as product test fixtures (the point of tracking this)

Several exit artifacts already flagged their probe assets as reusable when explain is built — these should become real test/ fixtures/anchors at that time, not be re-derived:

• Q14 mutation suite (-q14-recovery.ts) → recovery/diagnostics anchor tests (the fail-closed bare-word rule + defect-region contract).
• Q15 coordinate fixtures + reference mapper (-q15-coords.ts) → the coordinate-invariant anchor test when the coordinate mapper ships.
• Q17 adversarial generators (18 + 5 nesting shapes, -q17-stress.ts) → determinism / span-well-formedness / depth-bound property tests.
• Q0 gold set (-q0-gold-rev3.json) → if any IL projection is ever productized.
• Q13 highlight streams → the per-occurrence oracle for a span-vocabulary (Q12) build; large, CHR-derived, reproducible from -q13-recapture.ts against a CHR.

Decision needed

Pick a home so this survives a .scratch/ wipe without polluting mainline. Options (not mutually exclusive):

1. Keep in .scratch/, recover-by-hash. Cheapest. The exit artifacts already pin every file's sha256; re-runnable against the two CHRs. Risk: .scratch/ is explicitly ephemeral and uncommitted — one cleanup loses it.
2. Archive branch (e.g. lab/explain-185-archive) holding the framework verbatim, never merged. Durable, out of mainline history, git-recoverable. The 15 MB of streams would need .gitattributes/LFS or exclusion.
3. Promote-on-implement. Leave in .scratch/ now; when explain implementation starts, lift the four reusable suites above into test/ as fixtures/anchors and drop the rest. Cleanest long-term; depends on implementation actually starting.

Suggested: (1) now + (3) at implementation, with this issue as the durable index so nothing is orphaned. Streams (option-2 concern) stay CHR-reproducible and need not be committed.

Not blocking #185 closure or the ratification PR — this is bookkeeping for the reusable half.

[mobileskyfi]

Review follow-up from #187 (17aa534) turns this disposition into an explicit phase 0.5 — product contract fixtures gate before the offline core.

Recommended promotion set:

• representative frozen holdout + constructed corners from Q1–Q4, Q6, Q13, and Q16;
• all Q14 mutation operators and the no-fabrication/defect-region expectations;
• Q15 coordinate categories + six invariants, expressed against the production mapper;
• Q17 adversarial generators for determinism, span well-formedness, bounded depth/no-throw, and roughly-linear scaling;
• live Q6/Q8/Q10/Q11 anchors stay CHR integration tests in phase 2 rather than frozen raw captures.

The tests must target exported production interfaces and must not import or execute .scratch/ SUT/oracle code. The 76-file lab and the two large highlight streams remain out of mainline; only minimal reviewed inputs/expected outcomes move under test/.

This review did not find a reason to rerun or enlarge phase 0. Q5/Q7/Q9/Q12 are implementation/span-vocabulary refinements and should become focused phase-1/phase-4 tests as their surfaces harden.